Will 10 Gigabit Ethernet (10GbE) reach critical mass in the datacenter next year? The beginnings of a viable 10GbE ecosystem are now emerging and a number of analysts think 2008 may be a watershed year for this technology. Dell’Oro Group reports nearly a million 10GbE switch ports were shipped in 2007 and predicts 100 percent growth over the next two years. The Linley Group estimated 50 thousand 10GbE NICs were shipped in 2007 (twice as many as the previous year); however, 99 percent went into midrange Unix servers rather than x86-based systems. Although some formidable challenges remain, the industry may finally be ready to make the shift from Gigabit Ethernet (GigE) to 10GbE.
The key battleground for interconnect dominance will take place where high performance connectivity is in most demand: high end technical computing (science research, oil & gas, financial services, life sciences, digital animation, etc.) and high performance enterprise computing (Web 2.0, video editing & production, IPTV, real-time database applications, etc.). These types of applications have a critical need for more bandwidth and/or lower latency. With computing power being concentrated by multicore processors, blade servers and virtualization, server-to-server and server-to-storage communication has become a critical bottleneck.
While Fibre Channel is confined to storage networks, both InfiniBand and Ethernet have the ability to connect both storage and compute servers. Although InfiniBand is well established at the high end of HPC, the vast mid-market is still up for grabs, and will quickly respond to the best price/performance solution. The 10GbE switch and NIC vendors are counting on the ubiquity of Ethernet to make it a no-brainer as a unified server and storage fabric for the datacenter.
Analysts like Linley Group’s Bob Wheeler would support that contention. In a report about the benefits and challenges of 10GbE, Wheeler writes: “[T]the adoption of 10GbE is a certainty. The only questions are how quickly 10GbE will be adopted and to what extent it will displace alternative technologies such as Fibre Channel and InfiniBand.” In truth, adoption is long overdue; the aforementioned Linley report was written in January 2005.
InfiniBand vendors believe that the technical superiority and attractive price/performance of their technology have paved the way to the HPC mid-market as well as to the broader enterprise computing market, where highly dense and highly utilized hardware is creating the same demands on the interconnect as it has in traditional HPC. Over the past couple of years, InfiniBand has established itself as the de facto standard for high performance interconnects. More than half (58 percent) of the 100 fastest supercomputers now use InfiniBand, according the latest TOP500 list. Although the list is still dominated by Gigabit Ethernet overall, InfiniBand penetration has been doubling every year since 2005. (For one man’s view of the significance of InfiniBand’s rise in the TOP500, read Gilad Shainer’s article in this issue.)
By moving from Single Data Rate (10 Gbps) to Double Data Rate (20 Gbps) in 2007, InfiniBand has opened up a performance gap that 10GbE will be hard-pressed to fill. With Quad Data Rate InfiniBand (40 Gbps) products on the horizon, 10GbE will have to compete on something other than raw performance.
The big selling points for 10GbE are its position as the heir apparent of GigE and its ability to act as a unifying fabric for NAS, SANs, LANs, and cluster computing systems in the datacenter. NIC vendors like NetEffect, NetXen, and Chelsio Communications are offering 10GbE adapters with built-in support for RDMA (iWARP) to lighten the load on the CPU and achieve InfiniBand-like latency. Chelsio has been pushing its “unified wire” strategy based on its new 10GbE Terminator 3 ASIC. The chip has the ability to handle NIC, TOE, iSCSI and RDMA applications concurrently.
Up until this point, the 10GbE NICs and switches have been too expensive to be widely deployed in clustered systems. But with sub-$1000, RDMA-capable NICs starting to appear from vendors like Chelsio and $400-per-port switches from Arastra, that equation is changing. Arastra is using Fulcrum’s latest 10GbE switch silicon, which was designed to enable compute and storage clustering via Ethernet connectivity.
According to Chelsio CEO and President Kianoosh Naghshineh, once 10GbE NICs that support storage and server connectivity become standard on server motherboards, users will be faced with the decision to purchase additional InfiniBand and Fibre Channel HBA/HCAs, switches and gateways for clustering applications or to just use the omnipresence Ethernet NIC. He predicts 10GbE will have “an identical or better cost structure with InfiniBand by the end of next year .”
This year, Woven Systems, a startup switch vendor, set the stage for datacenter Ethernet when it released the EFX 1000, a 10GbE switch that performs active congestion management for lossless Ethernet. In tests at Sandia National Laboratories, researchers determined that the Woven switch actually outperformed an SDR InfiniBand setup when running a CBench performance test. The lab recently deployed the Woven switch for its 128-node “Talon” cluster.
“Ten gig is really ready for prime time in high performance computing and compute cluster interconnects,” said Derek Granath, Woven’s VP of marketing. “The ecosystem is mature.”
Finisar Corporation, an optical communications vendor, sees a big opportunity in fiber optical cables for Ethernet. The company recently announced its “Laserwire” optical cable assembly aimed at 10GbE connectivity. Unlike some of the other optical cable vendors, like Intel and Luxtera, which are more focused on InfiniBand connectivity, the Finisar assembly is implemented as a single 10 Gbps serial link optimized for 10GbE. It requires just 0.5 watts per optoelectrical transceiver at each end of the cable. The company is betting that the move to 10GbE in both HPC and the larger enterprise market is imminent and they want to be there to catch the wave.
To some extent, the OpenFabrics Enterprise Distribution (OFED) software stack is leveling the interconnect playing field by creating a unified platform for RDMA over Ethernet and InfiniBand. By making the software interoperable, users are free to take advantage of either technology without having to modify their code. The wild card here is Intel’s 10GbE approach, which rejects the RDMA/TOE model in favor of an I/O acceleration system that distributes network processing over the entire platform (the processor, chipset, network controller and software). Sun Microsystems announced its own 10GbE networking technology that offers a similar approach. At this point, the industry has not decided which approach to favor.
None of the 10GbE vendors think Ethernet is ready to replace InfiniBand at the high end of HPC. For those applications, InfiniBand’s bandwidth and latency cannot be matched by the current 10GbE solutions. In certain cases, application performance may be better with Ethernet solutions, based on more mature IP software stacks or the use of traffic routing in some of the newer switches. But for the most part, InfiniBand is about two years ahead of Ethernet in raw performance. InfiniBand switches and adapters that support QDR (40 Gbps) will begin to appear next year, while 40GbE is not expected to be ratified as a standard until 2010 (with volume product shipments years after that).
Rather than taking advantage of a deeply entrenched base, InfiniBand vendors are looking to work their way down from the TOP500 list into the volume market. Buoyed by a May 2007 InfiniBand study from IDC that projected a 40 percent compound annual growth rate in total factory revenue (InfiniBand switches and adapters) from 2006 to 2011, vendors are eyeing the broader enterprise market and moving to the next technology level.
Mellanox, the sole InfiniBand switch silicon vendor, recently announced its next generation InfiniScale IV chip, which supports QDR InfiniBand. The chip offers server-to-server and server-to-storage connections of 40 Gbps. (Because of inefficiencies and overhead in PCI Express 2.0, applications will realize only about 25 Gbps.) Latencies across a single switch chip are 60ns — less than a third of the latency of the top-of-the-line 10GbE switch silicon. The InfiniScale IV silicon supports 36 ports per switch, compared to the 24 ports supported by InfiniScale III technology and current 10GbE switch silicon. This works to reduce application-level latency for scaled-out clusters due to fewer switch hops. The latest chip also supports adaptive routing for those cases where static routing is not optimal, and congestion management to prevent the whole fabric from being overwhelmed with unusually heavy traffic. The new Mellanox switch chip is scheduled to be available in Q1 2008, and both QDR switches and adapters are expected to show up later in the year. QDR InfiniBand is likely to make its debut on the TOP500 list, perhaps by November 2008.
While the 10GbE proponents would love to get a system on the TOP500 list, the real battle is more likely to take place in the trenches of the datacenter, where sub-teraflop cluster systems are being sold in volume and are often connected in more loosely-coupled grids. Here 10GbE solutions may have an edge over InfiniBand since this type of installation is more likely to end up in an Ethernet-only facility. Also systems such as these often don’t need the maximum levels of connectivity performance offered by InfiniBand.
For their part, InfiniBand suppliers are hoping that increases in virtualization and compute and storage density continue to the point where connectivity demands outrun what 10GbE solutions can reasonably provide. In that scenario, InfiniBand is likely to become a much more mainstream solution. The next 12 months may provide some answers.